Views: 1
Artificial Intelligence is advancing at a spectacular pace. Every day, more powerful models, smarter assistants, and tools capable of generating text, images, videos, or code in seconds appear. But behind all this technological magic lies a less visible reality: the enormous energy consumption of the data centers that power modern AI.
Large LLM model training systems, like those used by the world's leading technology companies, are true industrial giants. We're talking about facilities with tens of thousands of specialized chips operating simultaneously and consuming colossal amounts of energy, in some cases approaching 5 gigawatts or even more.
And here lies one of the major current problems: connecting these mastodones to the electrical grid is no easy task. Electricity companies and governments must ensure that this enormous energy consumption does not affect the supply to other users. That's why permits, technical studies, and authorizations can take years. In many projects, the real bottleneck is no longer manufacturing chips or building servers, but rather securing sufficient access to the electrical grid.
However, there's a very interesting detail that has gone almost unnoticed for years. The world's electrical grids are designed to handle peak demand. This means that for much of the time they operate well below their actual capacity. In many cases, electrical infrastructures operate at 50% of their potential… or even less.
And that's where a brilliant idea emerges, seemingly simple yet effective. NVIDIA, understandably interested in fostering the development of more AI infrastructure, is collaborating with organizations like the Electric Power Research Institute, among others, to develop small data centers near existing electrical substations.
The logic is compelling. In the United States alone, there are more than 50,000 electrical substations. Many of them can handle between 5 and 40 MW, and even more in industrial areas. This opens the door to installing micro data centers of between 5 and 20 MW very close to these infrastructures, using available electrical capacity that is currently underutilized. This is also happening in the rest of the world.
Instead of waiting years to connect a mega data center to the grid, these small nodes could be deployed much faster by leveraging existing connections.
Furthermore, in many regions, substations are not too far apart. This would allow for the creation of distributed networks of interconnected data centers, sharing workloads and computing power.
The idea is somewhat reminiscent of how the internet works: instead of relying solely on a few gigantic centers, processing is distributed among multiple smaller nodes closer to the user.
And here another key factor emerges: AI inference. While training models requires enormous amounts of energy, inference—that is, using the already trained AI to answer questions, generate content, or perform tasks—consumes significantly less. This inference is precisely what millions of users need every day.
Therefore, these small, distributed data centers could become the perfect infrastructure for delivering fast, local, and efficient AI services.
There's another particularly clever aspect to this approach: during peak electricity demand, priority would still be given to the original users of the substation. The data centers would temporarily reduce their consumption to avoid straining the grid.
In other words, AI would adapt to the electrical grid… and not the other way around. This strategy could greatly accelerate the global deployment of AI services without requiring major expansions of national electrical infrastructure. Furthermore, it would reduce costs, bureaucratic delays, and technical complexity.
The NVIDIA-led team expects to have some of these first centers operational by the end of 2026. If the model works well, it seems highly likely that other countries will quickly adopt the idea.
Because, ultimately, great technological revolutions don't always stem from impossible inventions. Sometimes they appear simply by observing an existing system… and using it in a much smarter way.
EN 
ES